diff --git a/benchmarks/ios/jvm-vs-kotlin-native/src/commonMain/kotlin/MeasureComposable.kt b/benchmarks/ios/jvm-vs-kotlin-native/src/commonMain/kotlin/MeasureComposable.kt
index 6b11a8f842..98236a9de5 100644
--- a/benchmarks/ios/jvm-vs-kotlin-native/src/commonMain/kotlin/MeasureComposable.kt
+++ b/benchmarks/ios/jvm-vs-kotlin-native/src/commonMain/kotlin/MeasureComposable.kt
@@ -1,6 +1,9 @@
 import androidx.compose.runtime.Composable
-import androidx.compose.ui.ComposeScene
+import androidx.compose.ui.InternalComposeUiApi
+import androidx.compose.ui.graphics.asComposeCanvas
+import androidx.compose.ui.scene.MultiLayerComposeScene
 import androidx.compose.ui.unit.Constraints
+import androidx.compose.ui.unit.IntSize
 import kotlin.time.Duration
 import kotlin.time.ExperimentalTime
 
@@ -10,20 +13,20 @@ const val height = 480
 const val nanosPerSecond = 1E9.toLong()
 const val nanosPerFrame = (0.16 * nanosPerSecond).toLong()
 
-@OptIn(ExperimentalTime::class)
+@OptIn(ExperimentalTime::class, InternalComposeUiApi::class)
 fun measureComposable(
     frameCount: Int = 1000,
     content: @Composable () -> Unit
 ): Duration {
-    val scene = ComposeScene()
+    val scene = MultiLayerComposeScene(size = IntSize(width, height))
     try {
         scene.setContent(content)
-        scene.constraints = Constraints.fixed(width, height)
         val surface = org.jetbrains.skia.Surface.makeNull(width, height)
+        val canvas = surface.canvas.asComposeCanvas()
         return kotlin.time.measureTime {
             var nanoTime = 0L
             repeat(frameCount) {
-                scene.render(surface.canvas, nanoTime)
+                scene.render(canvas, nanoTime)
                 nanoTime += nanosPerFrame
             }
         }
diff --git a/benchmarks/kn-performance/src/commonMain/kotlin/Benchmarks.kt b/benchmarks/kn-performance/src/commonMain/kotlin/Benchmarks.kt
index cd2d006c3a..bf5f027254 100644
--- a/benchmarks/kn-performance/src/commonMain/kotlin/Benchmarks.kt
+++ b/benchmarks/kn-performance/src/commonMain/kotlin/Benchmarks.kt
@@ -127,9 +127,10 @@ fun runBenchmark(
     targetFps: Int,
     frameCount: Int,
     graphicsContext: GraphicsContext?,
+    warmupCount: Int = 100,
     content: @Composable () -> Unit
 ): BenchmarkStats {
-    val stats = measureComposable(frameCount, width, height, targetFps, graphicsContext, content).generateStats()
+    val stats = measureComposable(warmupCount, frameCount, width, height, targetFps, graphicsContext, content).generateStats()
 
     println(name)
     stats.prettyPrint()
diff --git a/benchmarks/kn-performance/src/commonMain/kotlin/MeasureComposable.kt b/benchmarks/kn-performance/src/commonMain/kotlin/MeasureComposable.kt
index f614668ea3..91d84814da 100644
--- a/benchmarks/kn-performance/src/commonMain/kotlin/MeasureComposable.kt
+++ b/benchmarks/kn-performance/src/commonMain/kotlin/MeasureComposable.kt
@@ -1,17 +1,17 @@
 import androidx.compose.runtime.Composable
-import androidx.compose.ui.ComposeScene
-import androidx.compose.ui.unit.Constraints
-import org.jetbrains.skia.DirectContext
+import androidx.compose.ui.InternalComposeUiApi
+import androidx.compose.ui.graphics.asComposeCanvas
+import androidx.compose.ui.scene.MultiLayerComposeScene
+import androidx.compose.ui.unit.IntSize
 import org.jetbrains.skia.Surface
 import kotlin.time.Duration
 import kotlin.time.Duration.Companion.nanoseconds
 import kotlin.time.ExperimentalTime
-import kotlin.time.measureTime
 import kotlinx.coroutines.*
+import kotlin.time.Duration.Companion.milliseconds
+import kotlin.time.TimeSource.Monotonic.markNow
 
 const val nanosPerSecond = 1E9.toLong()
-const val millisPerSecond = 1e3.toLong()
-const val nanosPerMillisecond = nanosPerSecond / millisPerSecond
 
 interface GraphicsContext {
     fun surface(width: Int, height: Int): Surface
@@ -19,51 +19,78 @@ interface GraphicsContext {
     suspend fun awaitGPUCompletion()
 }
 
-@OptIn(ExperimentalTime::class)
+expect fun runGC()
+
+suspend inline fun preciseDelay(duration: Duration) {
+    val liveDelay: Duration
+    if (duration.inWholeMilliseconds > 1) {
+        val delayMillis = duration.inWholeMilliseconds - 1
+        delay(delayMillis)
+        liveDelay = duration - delayMillis.milliseconds
+    } else {
+        liveDelay = duration
+    }
+    val start = markNow()
+    while (start.elapsedNow() < liveDelay){}
+}
+
+@OptIn(ExperimentalTime::class, InternalComposeUiApi::class)
 fun measureComposable(
-    frameCount: Int = 500,
+    warmupCount: Int,
+    frameCount: Int,
     width: Int,
     height: Int,
     targetFps: Int,
     graphicsContext: GraphicsContext?,
     content: @Composable () -> Unit
 ): BenchmarkResult = runBlocking {
-    val scene = ComposeScene()
+    val scene = MultiLayerComposeScene(size = IntSize(width, height))
     try {
         val nanosPerFrame = (1.0 / targetFps.toDouble() * nanosPerSecond).toLong()
         scene.setContent(content)
-        scene.constraints = Constraints.fixed(width, height)
         val surface = graphicsContext?.surface(width, height) ?: Surface.makeNull(width, height)
+        val canvas = surface.canvas.asComposeCanvas()
+
+        // warmup
+        repeat(warmupCount) {
+            scene.render(canvas, it * nanosPerFrame)
+        }
 
         val frames = MutableList(frameCount) {
             BenchmarkFrame(Duration.INFINITE, Duration.INFINITE)
         }
 
-        var nanoTime = 0L
+        var nextVSync = Duration.ZERO
+        var missedFrames = 0;
+
+        runGC()
+
+        val start = markNow()
 
         repeat(frameCount) {
-            val frameTime = measureTime {
-                val cpuTime = measureTime {
-                    scene.render(surface.canvas, nanoTime)
-                    surface.flushAndSubmit(false)
-                }
+            val frameStart = start + nextVSync
 
-                val gpuTime = measureTime {
-                    graphicsContext?.awaitGPUCompletion()
-                }
+            scene.render(canvas, nextVSync.inWholeNanoseconds)
+            surface.flushAndSubmit(false)
 
-                frames[it] = BenchmarkFrame(cpuTime, gpuTime)
-            }
+            val cpuTime = frameStart.elapsedNow()
 
-            val actualNanosPerFrame = frameTime.inWholeNanoseconds
-            val nanosUntilDeadline = nanosPerFrame - actualNanosPerFrame
+            graphicsContext?.awaitGPUCompletion()
 
-            // Emulate waiting for next vsync
-            if (nanosUntilDeadline > 0) {
-                delay(nanosUntilDeadline / nanosPerMillisecond)
-            }
+            val frameTime = frameStart.elapsedNow()
+
+            frames[it] = BenchmarkFrame(cpuTime, frameTime - cpuTime)
+
+            missedFrames += (frameTime.inWholeNanoseconds / nanosPerFrame).toInt()
 
-            nanoTime += maxOf(actualNanosPerFrame, nanosPerFrame)
+            nextVSync = ((it + 1 + missedFrames) * nanosPerFrame).nanoseconds
+
+            val timeUntilNextVSync = nextVSync - start.elapsedNow()
+
+            if (timeUntilNextVSync > Duration.ZERO) {
+                // Emulate waiting for next vsync
+                preciseDelay(timeUntilNextVSync)
+            }
         }
 
         BenchmarkResult(
diff --git a/benchmarks/kn-performance/src/desktopMain/kotlin/runGC.jvm.kt b/benchmarks/kn-performance/src/desktopMain/kotlin/runGC.jvm.kt
new file mode 100644
index 0000000000..3095a116b1
--- /dev/null
+++ b/benchmarks/kn-performance/src/desktopMain/kotlin/runGC.jvm.kt
@@ -0,0 +1,3 @@
+actual fun runGC() {
+    System.gc()
+}
\ No newline at end of file
diff --git a/benchmarks/kn-performance/src/macosMain/kotlin/runGC.native.kt b/benchmarks/kn-performance/src/macosMain/kotlin/runGC.native.kt
new file mode 100644
index 0000000000..7b499db732
--- /dev/null
+++ b/benchmarks/kn-performance/src/macosMain/kotlin/runGC.native.kt
@@ -0,0 +1,7 @@
+import kotlin.native.runtime.GC
+import kotlin.native.runtime.NativeRuntimeApi
+
+@OptIn(NativeRuntimeApi::class)
+actual fun runGC() {
+    GC.collect()
+}
\ No newline at end of file